Your next voice assistant might live on your laptop instead of in a data center. Izwi 0.1.0, a free local voice AI desktop app, just reached its first stable release after months of public betas. It's released under Apache-2.0, an open-source license that lets you use, modify, and build on the code. Izwi handles the voice side of local AI: it listens with speech recognition, thinks with a language model, and talks back with text-to-speech. It also comes as a web UI, a command-line tool, and a local server, so it's more than a single desktop window.
What happened
Izwi isn't one AI model. It's a control room for several local models that handle different jobs. The core trick is a real-time voice loop: local ASR turns your speech into text, an LLM writes a reply, and TTS reads that reply out loud. ASR means automatic speech recognition, or speech-to-text. An LLM is the chat brain. TTS is text-to-speech, the part that gives the machine a voice.
The local voice loop
That loop is the headline feature. You speak, Izwi transcribes, a chat model answers, and the app speaks the answer back. The same app also does TTS on its own, voice cloning from a short sample, and voice design from a text description. Voice cloning means giving the app a small voice sample so it can create a synthetic version. Voice design means describing a voice in words and letting the model build it.
There's also a Studio for long-form audio like audiobooks, plus transcription with speaker diarization and forced alignment. Diarization answers the question of who said what. Forced alignment lines up the transcript with the exact moments in the audio, so you can jump to the right spot instead of scrubbing through a recording.
Models and the API
For developers, the local server exposes an OpenAI-compatible /v1 API for chat completions, audio/speech, and audio/transcriptions. In plain English, code written for OpenAI's API can point at localhost instead. That's a big deal for anyone who already has a script or prototype and wants to test it with local models.
Inference runs natively in Rust on Candle, with the project's own Metal and CUDA work, so there's no Python environment to manage. Candle is a Rust framework for machine learning. Metal is Apple's GPU layer, and CUDA is Nvidia's GPU layer. The practical win is fewer setup headaches for people who don't want to babysit a Python stack.
The supported model list is broad. TTS options include Qwen3-TTS, Kokoro-82M, Fish Audio S2 Pro, VibeVoice, Voxtral TTS, and LFM2.5 Audio. ASR options include Parakeet TDT v3, Whisper large-v3-turbo, Qwen3-ASR, Granite Speech 4.1, Nemotron ASR, and Voxtral Mini. Diarization uses Sortformer and Nemotron 3 diarization, while chat runs on Qwen chat models and LFM2.5, so you can pair a heavier chat brain with a lighter voice model.
What it means for you
This isn't just a developer toy. It changes what ordinary people can do with voice AI on a normal computer, because the voice pipeline is local and the app is free to use under Apache-2.0. Here's what that looks like in real life.
At home
You can turn family voice notes, interviews, or school recordings into text without sending personal audio to a random website. Transcribe a recording, then ask a chat model to turn the notes into a tidy summary. It's handy when privacy matters and you'd rather keep the files on your own machine.
At work
Meeting notes get much easier when the transcript labels who said what. Speaker diarization separates the voices, and forced alignment helps you find the exact part of the audio behind a quote. Instead of replaying a whole call, you can jump to the moment that matters.
For business
Small teams can test a voice assistant before paying for cloud infrastructure. Point a prototype at the local /v1 API, use chat completions and audio/speech, and see how the flow feels. Because Izwi is Apache-2.0, you can experiment without a license fee.
If you want a wider set of free AI tools to compare, MyKreaTool is a useful place to browse. It's a natural companion when you're testing local AI and want more free options.
For study
Lecture recordings and interview clips become searchable text. You can transcribe them, then use forced alignment to connect the transcript to exact timestamps. That makes revision faster because you can go straight to the section you need instead of listening from the start.
For creativity
Podcasters and indie authors can design a voice from a text description or clone a voice from a short sample, then use the Studio for long-form audio like an audiobook draft. Voice design is especially fun: you describe the voice you want in words, and the model builds it. You still need to respect consent and rights when you clone a voice.
For income
Freelancers who build voice demos can spin up a local prototype and show clients a working flow. You're not promising studio quality. You're cutting the time between idea and demo. The OpenAI-compatible API means existing scripts can point at localhost instead of a paid endpoint, which makes paid experimentation less painful.
How to try it right now
Start with the free option: Izwi 0.1.0 itself. It's free and Apache-2.0, so there's no paid license to test. Here's a simple path.
1. Find the official Izwi 0.1.0 release from the project's own page. Avoid reposted binaries.
2. Download the desktop build for your system. You can also use the web UI, CLI, or local server if that fits your workflow.
3. Open Izwi and choose models from the supported list. For a quick TTS test, pick a listed model such as Kokoro-82M. For speech recognition, try Whisper large-v3-turbo or Parakeet TDT v3.
4. Run a short voice conversation: speak into your mic, let local ASR turn it into text, let a chat model such as a Qwen model respond, and let TTS read the answer.
5. Open transcription and test a clip with two speakers. Turn on diarization to see who said what, and use forced alignment to line up words with the audio.
6. If you code, start the local server and point an OpenAI-compatible client at localhost /v1. The endpoints include chat completions, audio/speech, and audio/transcriptions.
7. For long-form audio, open the Studio and test an audiobook-style draft.
Upsides and what changes
Your voice workflow becomes local
The pitch is simple: the voice pipeline runs on your machine, not through a cloud API. That gives you more control over your audio and your models. You choose from the supported list, including small options like Kokoro-82M and larger ASR options like Whisper large-v3-turbo. You're not locked into one vendor's voice stack.
Builders get a familiar API
Existing OpenAI-style code can be reused. You can swap a cloud endpoint for localhost and test chat, speech, and transcription through the same /v1 shape. That lowers the cost of trying a local voice agent. The Rust and Candle setup also means there's no Python environment to manage, which is a real quality-of-life change for people who've spent hours fixing package conflicts.
The biggest shift is that voice AI becomes a tool you install, not a service you rent. You can test voice cloning, voice design, and long-form audio without building a cloud pipeline first.
Limitations
Izwi is a 0.1.0 release, so treat it as an early stable build, not a finished consumer product. Your hardware and model choice will decide how fast it feels and how good it sounds, and you'll still need to download models and manage the local setup yourself. Voice cloning and voice design also raise consent and copyright issues, so get permission before using someone else's voice. And because it's local, you're the one responsible for updates, troubleshooting, and keeping your setup running.
Conclusion
Izwi 0.1.0 makes local voice AI feel like a real desktop tool: speech recognition, chat, text-to-speech, voice cloning, voice design, transcription, and an OpenAI-compatible API in one free Apache-2.0 package. It won't replace every cloud voice service, but it gives you a practical way to test the local-first approach without a paid license. Your action for today: download Izwi 0.1.0 and run a short listen-transcribe-speak test on your own machine.
Comments 0